1. Field of the Invention
The present invention relates to an optical information processing apparatus and method for an image sensing device of an industrial robot or the like which carries out filtering of an input image in a spatial frequency range, image processing such as feature extraction or the like, or discriminating an input pattern coincident with a specific reference pattern among a plurality of input images.
2. Description of the Prior Art
A conventional optical information processing apparatus of this type is disclosed, for example in Japanese Patent laid-open publication No. H2-132412.
FIG. 7 shows a fundamental composition thereof.
In this apparatus, when an object is picked up by a TV camera 401, the image thereof is displayed on a first liquid crystal display 402 which is irradiated by a coherent light emitted from a laser diode 403 and collimated by a collimator lens 404. Since the first liquid crystal display 402 is arranged in the front focal plane of a first lens 405, a Fourier transformed image of the object is formed on a second liquid crystal displayed 406 arranged in the rear local plane of the first lens 405. At this time, a Fourier transformed image of a specific reference pattern is simultaneously displayed as an optical filter on the second liquid crystal display 406 in a form of a Fourier transformation computer generated hologram by modulating spatially the transmittance of each pixel thereof using data related to the specific reference pattern which is memorized in a ROM 407. Accordingly, the Fourier transformed images of the object and the specific reference image are superposed on the second liquid crystal display 406.
Since the second liquid crystal display is arranged in the front plane of a second lens 408, these Fourier transformed images are optically Fourier transformed by the second lens 408. If both Fourier transformed images on the second display 406 coincide with each other, a bright point is generated on a rear focal plane of the second lens 408 and is detected by a photodetector 409. Thus, the object is discriminated by detecting the bright point.
However, in the conventional apparatus mentioned above, it is impossible to perform an exact pattern matching in a case in which an object is varied in the scale thereof or rotated since correlation factors between the images of the object and the reference pattern are varied thereby.
In order to solve this problem, D. Casasent et. al., proposed to perform a pattern matching between the image of the object and the reference pattern after executing a coordinate transformation of the image of the object which is invariant to variation in the scale thereof or rotation thereof [See D. Casasent et. al., Appl. Opt. 26,938 (1987)].
However, according to the composition proposed by D. Casasent et. al., a plurality of input patterns have to be interchanged in turn and, also, it becomes necessary to interchange a plurality of phase filters for coordinate transformation in turn upon performing plural coordinate transformations. However, since it becomes necessary to position each phase filter at an extremely high accuracy, real-time processing for the coordinate transformation of the input pattern is impossible, resulting in a lack of flexibility.
Furthermore, the apparatus of this type has a disadvantage in that it is difficult to recognize an object exactly when the object is moved parallel to the origin of a logarithmic polar coordinate since scale and rotation invariance is obtained only in a case that the center of the object coincides with the origin of the above coordinate.
This will be explained below using FIGS. 8 and 9.
In FIG. 8(b) shows a pattern obtained by rotating the pattern of FIG. 8(a) by 90.degree. about the coordinate origin, FIG. 8(c) shows a pattern obtained by magnifying the pattern of FIG. 8(a) by k-times and FIG. 8(d) shows a pattern obtained by shifting the pattern of FIG. 8(a) in the x-direction by a distance m. FIGS. 9(a), 9(b), 9(c) and 9(d) show patterns obtained by logarithmic polar coordinate transforming the patterns of FIGS. 8(a), 8(b), 8(c) and 8(d). As is apparent, the patterns of FIG. 9(b) and 9(c) are obtained by parallel-shifting the pattern of FIG. 9(a). Since the shift invariance is maintained in the pattern matching with use of a Fourier transforming optical system, patterns of FIGS. 8(a), 8(b) and 8(c) are recognized same and, accordingly, the scale and rotation invariance is maintained at the origin. However, the pattern FIG. 9(d) being the pattern obtained from the pattern of FIG. 8(d ) is quite different from the pattern of FIG. 9(a) and, accordingly, the former is not recognized as the same as the latter.